Replit Makes OpenAI the Default Behind Replit Agent. What the CEO Said, What Is Disputed, and What Enterprise Admins Can Actually Pin

Most people never change the default model in a coding tool. That is why a change to Replit's default matters more than the fact that Claude is still in the menu.

What Masad said, in order

• On stage at Disrupt (16 October): Replit is "no longer working with Anthropic." Masad said the decision was based on "cost, partnership, and performance," and that on Replit's internal and public benchmarks OpenAI's Sol and Astra "perform better on building vibe-coded applications from scratch." He also said Replit had until recently been a top-five Anthropic customer.
• After Disrupt: customers can still use Anthropic models, but OpenAI's models are now the default.
• Statement to TechCrunch over the weekend: "Replit remains model agnostic." Users can choose Claude or other supported models, and "OpenAI is now our preferred partner for powering Replit Agent."

What is disputed

TechCrunch reports that a person familiar with the Anthropic and Replit partnership, who asked not to be named, said it was still in place. TechCrunch also points to a late-September Replit blog post that named Claude Fable 5 and Fable 5.1 as models running in Replit Agent's production environment. Neither Anthropic nor OpenAI has commented on the record in the reports we read, so whether the commercial partnership has ended is still an open question. The product change is not: by the CEO's own statement, OpenAI is the default.

The move follows Replit's August launch of Free Mode, which its own blog says is powered by OpenAI's GPT-5.6 Luna and lets Core subscribers on the $20 plan do everyday Agent tasks without spending credits.

Why customers get no notice

Rajesh Beri's analysis of Replit's legal documents explains the gap. Replit's subprocessor list, last updated in August 2026, already names OpenAI, Anthropic and Google for AI models. Its data processing agreement, dated 1 April 2026, gives Replit general authorisation to use subprocessors and lets customers object only to "a change of a Subprocessor." Shifting more traffic between two providers that were both already listed is hard to argue is such a change, so nothing in the standard paper obliges Replit to tell you first.

What enterprise admins can and cannot control

• You cannot remove OpenAI. Replit's enterprise model controls say: "At least one model from OpenAI, Anthropic, and Google must remain enabled."
• You can narrow the list. Each provider can be set to Enable all, Enable selected or Disabled. The strongest pin is one OpenAI model, one Google model, and the Claude model your team has tested.
• Routing is on by default. Since 28 August, Intelligent Model Routing with Auto on has been the default Agent mode for Enterprise Workspaces, so a task can still be sent to the OpenAI or Google model you left enabled. Beri found no admin setting that forces Auto off for everyone, and Replit has not said how "preferred partner" changes routing weights.
• These are not spending caps. Replit's documentation says so directly.

Why Replit is making the change now

Masad told the Disrupt audience that enterprise contracts were about 2% of revenue last year and about 15% this year, and that more than half of revenue is now business-to-business. Backstage he said he expects enterprise and government customers to pass half of revenue in 2027. Those are the customers whose privacy reviews assumed a particular model provider.

What this means if you build on Replit

• If you have never touched the model selector, your Agent work now runs on OpenAI models by default. Re-test apps you built while Claude was the default before trusting large changes.
• If you are on an enterprise plan, review the provider settings in the admin console and ask your account team in writing about data retention with OpenAI and how Auto mode now routes tasks.
• Treat Masad's benchmark claim as a vendor claim. It measures building new apps from scratch, which may not match extending an app you already have.